Qinan Huang黄祺楠 Machine learning for chemistry
email GitHub X LinkedIn

Machine learning for chemistry

Qinan Huang黄祺楠

PhD candidate @Amanchukwu Lab
Pritzker School of Molecular Engineering
The University of Chicago
qinanh@uchicago.edu
A battery cell — ions moving through the electrolyte − + anode cathode electrolyte
scroll down to continue

about

I'm a PhD candidate in the Amanchukwu Lab at the Pritzker School of Molecular Engineering, The University of Chicago, advised by Chibueze Amanchukwu. Before that a BS in chemistry at Hunan University (2024), then an MS in molecular engineering at UChicago (2026).

I work on machine learning for chemistry — models that predict how molecules behave, and the methods that make those predictions reliable enough to build on.

The questions I keep coming back to: predicting molecular properties from structure alone; what statistical mechanics can teach a learned model about collective behaviour; and how to describe solution systems, where the chemistry that matters lives in the crowd, not in any single molecule. The thread through all of it: turning chemical intuition into something a machine can check, not just repeat.

an electrolyte box
an electrolyte box
Li⁺ first shell
Li⁺ first shell
EC π* LUMO
EC π* LUMO
a learned Li field
a learned Li field

news

Aug ’26
Talk at ACS Fall 2026 — Teaching Language Models Electrolyte Physics with Verifiable Rewards.
Aug ’26
Released xyzrender — molecules and orbitals to transparent SVG in one command.

what I work on

three threads — click one to open it

Language models that reason about physical processes

I post-train language models with reinforcement learning whose reward comes from physics rather than labels: the model reasons through a mechanism or a structure, and simulation decides whether the reasoning holds. Pointed at electrolytes, that turns a months-long search into a shortlist.

→ the grader is physics, not a label

Predicting the properties of mixtures, as mixtures

Almost everything worth predicting is a mixture, and models built on pure components handle them badly — what matters is how the components sit next to each other. I learn a representation with composition in the input, so non-ideal, concentration-dependent behaviour is something the model can express rather than something bolted on.

→ composition in, property out

Designing solvation structure instead of guessing formulations

An average coordination number hides the shape of the crowd, and the shape is what governs transport and stability. I train generative models anchored in the solution's statistical mechanics, which invert the search: describe the solvation structure you want, and the model proposes the solvents and salts that produce it.

→ design the shell, not the solvent

publications

selected · all on Scholar

code

开源的东西
xyzrender

XYZ / PDB / SMILES / cube → publication-quality transparent SVG, PNG and GIF. Molecular orbitals, MD boxes, solvation shells.
simulation_workflow

Li-metal electrolyte MD + screening pipeline. LigParGen → xTB descriptors → GROMACS → transport and RDF analysis.
ElectraG

Retrieval-augmented prediction of Coulombic efficiency from electrolyte composition, grounded in 153 papers.